I have to admit I consider myself an image geometry (and robotics) person first, but it’s clear that the application of machine vision to do identification, classification, and analysis — particularly given the amazing results shown by “deep learning” — has made this a hot topic.

I’ve found myself side-loaded into this topic from a couple of different angles. First, I think machine learning approaches, particularly deep NNs, are here to stay, better to be more conversant with them than less. Future research will look at combining model-based methods with learned methods for a variety of robotics topics: perception, control, multi-vehicle coordination, communication, etc.

Second (as a more specific follow-up to the above), I think image analysis can be a powerful complement to geometric approaches, particularly for applying meaning to what’s in the image. In my case, I understand how SLAM-like techniques can be used to produce a volumetric representation of an underwater environment, then a NN-like approach could be used to segment out fauna, identify features of interest, etc.

Third, the CamHD is challenging but only marginally interesting from a geometric point of view, but is a rich data set for image analysis. In fact, our goals for “deriving science” from CamHD is pure image analysis, using both model-based methods and potentially trained methods to convert images to data.

I’ve outlined some of the more specific research interests on the CamHD page.

Sample imagery from CamHD (Image courtesy NSF OOI)

Finally, I was able to use benthic visual surveys from the RSN construction to try my hand at using a CNN for animal identification. This has given me entry to the greater world of folks doing underwater image analysis for fisheries. I personally think this is an incredibly hard problem, but I’m looking for ways I can either contribute or help form a connection between those on the water and those on campus who might be able to help.

Sample imagery from Benthic survey (Image courtesy OOI/Ropos, Deb Kelley and Katie Bigham)

All video analytics posts: